3 ms·

> we can (or soon will be) able to read thought patterns via LLMs analyzing brain waves source ? who says that the brain waves we can measure hold information

> we can (or soon will be) able to read thought patterns via LLMs analyzing brain waves source ? who says that the brain waves we can measure hold information like that in a useful level of granularity ? you can have something that distinguishes "angry" from "happy" etc, sure, but thought patterns ?

chriskanan 3y ago
Here are a couple papers related to this topic, but the primary focus has been on showing that LLM embeddings correlate with neural activity better than alternatives. But given one has neural recordings, one could train a model to take them as input and output what was heard. How well that would work, I think, is an open question. https://www.biorxiv.org/content/10.1101/2023.06.27.546708v1 https://www.biorxiv.org/content/10.1101/2023.06.27.546708v1 https://www.biorxiv.org/content/10.1101/2022.07.11.499562v3 https://www.biorxiv.org/content/10.1101/2022.07.11.499562v3
- nickphx 3y ago
  Wouldn't the model need to be trained on some valid data before being used to guess/approximate output from input?
  - bgnn 3y ago
    yeah. yet people believe AI can do new things without training.
caycep 3y ago
It's actually a pretty big field in neurophysiology along several levels, and not just w/ LLMs, ppl are trying to throw more basic EE DSP at it. Most analyses can do simple things - i.e. differntiator/discriminator type things, but who's to say if recording technology gets better, more complicated thought patterns can't be interpreted LFP - "macro recordings" from larger electrical contacts like EEG, DBS leads, ECoG, i.e. waves from large areas of the brain single unit - smaller electrodes thought to capture recordings from 1 or a few neurons. There's actually a ton of stuff looking at differential responses of single unit brain waves against a lot of different psychological stimuli. NEURIPS had a couple of presentations on LLMs interpreting stuff from fMRI - maybe not "brain waves" technically speaking but it may be approaching "thought patterns".

chriskanan 3y ago

Here are a couple papers related to this topic, but the primary focus has been on showing that LLM embeddings correlate with neural activity better than alternatives. But given one has neural recordings, one could train a model to take them as input and output what was heard. How well that would work, I think, is an open question. https://www.biorxiv.org/content/10.1101/2023.06.27.546708v1 https://www.biorxiv.org/content/10.1101/2023.06.27.546708v1 https://www.biorxiv.org/content/10.1101/2022.07.11.499562v3 https://www.biorxiv.org/content/10.1101/2022.07.11.499562v3

nickphx 3y ago
Wouldn't the model need to be trained on some valid data before being used to guess/approximate output from input?
- bgnn 3y ago
  yeah. yet people believe AI can do new things without training.

caycep 3y ago

It's actually a pretty big field in neurophysiology along several levels, and not just w/ LLMs, ppl are trying to throw more basic EE DSP at it. Most analyses can do simple things - i.e. differntiator/discriminator type things, but who's to say if recording technology gets better, more complicated thought patterns can't be interpreted LFP - "macro recordings" from larger electrical contacts like EEG, DBS leads, ECoG, i.e. waves from large areas of the brain single unit - smaller electrodes thought to capture recordings from 1 or a few neurons. There's actually a ton of stuff looking at differential responses of single unit brain waves against a lot of different psychological stimuli. NEURIPS had a couple of presentations on LLMs interpreting stuff from fMRI - maybe not "brain waves" technically speaking but it may be approaching "thought patterns".